Before a percent becomes a fate
What Percentage of Startups Fail? What the Data Actually Says
“What percentage of startups fail?” is two questions pretending to be one: which population, and what counts as failure. This page puts a government survival table next to a well-known venture-backed definitions ladder, then asks you to write your own failure definition and kill line before you treat any of those numbers as weather.
Last updated: October 7, 2026
Direct answer
What percentage of startups fail?
In U.S. government data on new private-sector employer establishments, about 1 in 5 is gone within a year, roughly half survive five years, and about a third survive ten (BLS Business Employment Dynamics, Table 7). Venture-backed startups are a different population, and their “failure rate” ranges from 30–40% to over 90% depending on whether failure means liquidation, not returning capital, or missing projections (Shikhar Ghosh, as reported by HBS Working Knowledge). The number moves with the population and with the definition. A base rate is an outside view of the average new business, not a forecast for yours.
Key takeaways
What to remember
- Name the population before you quote a number: employer establishments, firms, venture-backed companies, and a side project with no payroll are not the same census.
- Name the failure definition before you quote a number: closed, liquidated, did not return capital, or missed the founder’s own plan are different events.
- “90% of startups fail” is usually quoted without a primary source. This page does not attribute that slogan, and it does not claim a named study has debunked it.
- A base rate is a prior about the average new business. It is not your fate, and a Yibud score is not a survival probability.
- Before you start, write your failure definition, a kill line and date, and the riskiest assumption you will test first. That is Yibud’s suggested template, not a law.
Why the slogan fails
“90% of startups fail” is usually quoted without a primary source
I have watched founders treat that sentence as if it were a weather forecast for their Tuesday. It is usually repeated without a dataset you can open. This page will not pin the slogan on a named author, and it will not claim a named study has knocked it down. What you can do instead is pick a population you can actually count, pick a definition of failure you would honor, and then read a source that measured that pair. The BLS table below measures whether a new U.S. private-sector employer establishment still has positive employment in a later March. The Ghosh ladder, reported by HBS Working Knowledge and the Wall Street Journal, measures venture-backed outcomes under four different definitions. Those are not the same number, and neither one is a verdict on a solo founder with no payroll.
Vocabulary
Define the population — and failure — before the percent
Most fights about “the” failure rate are fights about nouns. These six are the nouns this page uses on purpose.
1. Employer establishment — one location with payroll
In the BLS Business Employment Dynamics table on this page, an establishment is a single physical location. The data come from state unemployment-insurance records (QCEW): a census of employers, not a survey. Establishments with zero employment, and self-employed people without payroll, are excluded. A solo founder with no payroll is not in this table.
2. Firm — the company, which may own many establishments
A chain that opens a new branch creates a new establishment. One firm can have many. Quoting an establishment survival rate as if it were a firm failure rate mixes two units. This page does not convert one into the other.
3. Venture-backed startup — a funded firm, not the BLS census
The Ghosh figures reported below describe U.S. companies that received venture funding — in the Wall Street Journal write-up, generally at least $1 million, from 2004 through 2010, more than 2,000 companies. That is a different population from “every new employer establishment” and from a nights-and-weekends side project.
4. Survival rate — still operating with positive employment
BLS Table 7’s survival rate is the share of an opening-year cohort still operating — meaning positive employment — in the March of a later year. The opening year is the year ended March of the year shown. The extract on this page runs through March 2025. Survival is not success. A stagnant shop with one employee still counts as operating.
5. Failure — closed, liquidated, unpaid investors, or a missed plan
A closure is not always a failure (the founder sold, took a job, or hit a pre-written kill line). Liquidating assets so investors lose most or all of their money is one definition. Not returning investors’ capital is another. Missing a projected return, or missing a projection the founder stated, are two more. Write which one you mean before you quote a percent.
6. Base rate — the outside view, not a personal forecast
A base rate is what usually happens in a named population under a named definition — the outside view. It is a prior you can update with evidence about your specific bet. It is not a probability that you will fail, and it is not what a Yibud score measures.
The government table
BLS Table 7 — survival of new private-sector employer establishments
U.S. Bureau of Labor Statistics, Business Employment Dynamics, “Table 7. Survival of private sector establishments by opening year,” Total private. Survival rate = percent of the opening-year cohort still operating (positive employment) in the March of each later year. Opening year ended March of the year shown. Data run through March 2025. Use these values as printed.
| Opening year | New establishments | After 1 year | After 5 years | After 10 years |
|---|---|---|---|---|
| 1994 | 569,387 | 79.6% (Mar 1995) | 49.6% (Mar 1999) | 33.6% (Mar 2004) |
| 2014 | 652,518 | 79.7% | 50.8% (Mar 2019) | 34.9% (Mar 2024) |
| 2015 | 677,876 | 79.6% | 50.2% (Mar 2020) | 34.7% (Mar 2025) |
| 2020 | 767,573 | 80.9% | 51.4% (Mar 2025) | — |
| 2024 | 988,310 | 77.9% (Mar 2025) | — | — |
The 1994 cohort’s later readings, same table: 20.3% after 20 years (Mar 2014); 12.6% still operating in Mar 2025. Younger cohorts have not had twenty years yet. This page does not invent industry splits from Table 7.
In plain language, across these cohorts: roughly 20% closed within a year, roughly half closed within five years, and roughly two-thirds closed within ten. “Closed” here means no longer operating with positive employment. It does not mean unprofitable, and it does not mean the founder failed.
Stability
The curve barely moves across thirty years — including recessions
The teaching point in Table 7 is not a single lucky cohort. It is how tight the ranges stay. A founder who waits for “a better year” is usually waiting for a curve that does not move much.
1-year survival, opening years 1994–2024
Lowest 75.2% (2008 cohort, measured Mar 2009). Highest 80.9% (2020 cohort). Most of the table sits near 80%. The 2024 cohort’s 77.9% at one year (Mar 2025) is inside that band.
5-year survival, opening years 1994–2020
Lowest 45.4% (2006 cohort, Mar 2011). Highest 51.9% (2018 cohort, Mar 2023). Roughly half is not a slogan. It is where the five-year column lives.
10-year survival, opening years 1994–2015
Lowest 32.4% (2001 cohort, Mar 2011). Highest 35.3% (2010 cohort, Mar 2020). About a third still operating at ten years is a stable reading, not a one-off.
What the table is not
Four BLS caveats you have to keep when you quote these percents
These limits come from the BLS Business Employment Dynamics FAQ. Skip them and you are quoting a different study than the one you linked.
Employers only — no payroll, not in the data
The figures come from state unemployment-insurance administrative records (QCEW). That is a census, not a survey. It covers employers. Establishments with zero employment and self-employed people without payroll are excluded. If you have not hired, you are not in Table 7.
An establishment is not a firm
One location, not one company. A multi-site firm that opens a branch is a new establishment. A firm that shuts one store and keeps the others is not “a failed startup” in this table.
U.S. private sector only
These are United States private-sector establishments. They are not a world census, not a public-sector census, and not a China / EU small-business census. Do not import the percent into another country and call it the same fact.
Operating means positive employment — not profit, not growth
A location that still has payroll in March counts as surviving. The table is silent on whether the owner took a salary, whether revenue grew, or whether anyone was happy. Survival is not success. A closure is not always a failure.
A different population
Venture-backed “failure” moves when you change the definition
Shikhar Ghosh, then an HBS senior lecturer, gave a definitions ladder that the press repeated. Label these as one researcher’s estimates reported by HBS Working Knowledge (7 March 2011) and the Wall Street Journal (Deborah Gage, 20 September 2012) — not an official dataset, and not the BLS table.
Liquidation — 30–40%
If failure means liquidating all assets with investors losing most or all of their money, Ghosh’s figure is 30–40% (HBS Working Knowledge, 2011). That is the narrowest definition on this ladder.
Did not return capital — about three-quarters
Gage’s Wall Street Journal piece (20 September 2012) reports Ghosh’s finding that about three-quarters of U.S. venture-backed firms do not return investors’ capital, based on more than 2,000 companies that received venture funding, generally at least $1 million, from 2004 through 2010. Same article: the National Venture Capital Association estimated 25% to 30% of venture-backed businesses fail — a different definition, printed next to Ghosh’s, not a reconciliation.
Missed the projected return — 70–80%
If failure means not seeing the projected return on investment, Ghosh’s figure is 70–80% (HBS Working Knowledge, 2011). Most of the companies in that band still exist. They just did not deliver the return that was modeled.
Missed a stated projection — 90–95%
If failure means declaring a projection and falling short, Ghosh’s figure is 90–95% (HBS Working Knowledge, 2011). That is how a “90%” sentence can be true inside one definition and misleading as a slogan about every new business.
When the death happens — first four years vs after year four
The same Journal article reports Ghosh’s timing claim: non-venture-backed companies fail more often in the first four years; venture-backed ones tend to fail after year four, once investors stop injecting capital. That is a reported estimate about timing, not a BLS finding, and not a reason to skip a kill line in year one.
Reasons are not rates
A list of why companies died is not the probability you will die
CB Insights’ 2026 analysis of 431 venture-backed companies that shut down since 2023 — 385 with identifiable reasons — is already cited elsewhere on Yibud as pattern frequency among shutdowns, not as a probability of shutting down. “Ran out of capital” leads their list at 70 percent; they call it often the final cause of death. Among the 385, more telling patterns include poor product-market fit (43 percent), bad timing (29 percent), and unsustainable unit economics (19 percent). Many shutdowns cited more than one reason, so those shares can add to more than 100 percent. Use those figures as “among companies that already died, these reasons showed up this often.” Do not turn 43 percent into “you have a 43 percent chance of no market need.” This page invents no other CB Insights percentages.
How to use a base rate
Write the outside view, then write your own lines
Do this before the repo exists. The BLS table and the Ghosh ladder are priors. They do not replace a definition you would actually honor. Yibud’s template below is a suggested sitting, not a statistical law.
- 1
Name the population you actually belong to
Employer establishment, multi-site firm, venture-backed company, or a no-payroll side project. If you have not hired, Table 7 is a neighbor’s census, not yours. If you have not raised venture money, Ghosh’s ladder is a neighbor’s census, not yours. Write the noun.
- 2
Write your failure definition in your terms
Money, time, or a signal a stranger could log. “I will have spent six months and still have fewer than three paying customers in the named segment” is a definition. “It didn’t work out” is a mood. Closure, liquidation, and “missed my own plan” are allowed. Pick one.
- 3
Write a kill line and a date
A threshold plus a calendar day. The kill-criteria page is the full Stop contract (signal, threshold, window, named action). This step is why that contract needs a definition of failure before hope edits the bar.
- 4
Name the riskiest assumption you will test first
The one whose failure would make the rest of the plan irrelevant. The critical-assumption page is the week. The pre-mortem page is how you invent the failure history if you do not yet have a ranked list. Test that assumption before you decorate the product.
- 5
Keep the base rate as a prior — then look at your evidence
“About half of new U.S. employer establishments are gone by year five” is a reason to write the kill line now, not a reason to quit this afternoon. Update the prior with the cheapest evidence you can actually collect. Do not update it with a mid-band Yibud score and call that survival.
What you leave with
The failure-definition sheet (copy this)
If the sitting produced a vibe, you did not finish. Fill the brackets. Leave nothing as “startups,” “success,” or “we’ll know.”
The four fields
My population: [employer establishment / firm / venture-backed / no-payroll side project — pick one]. My failure definition: [money, time, or a signal a stranger could log]. My kill line + date: [threshold] by [calendar day] → [Kill / Pivot / Continue]. Riskiest assumption first: [the one claim whose failure would make the rest irrelevant].
If any bracket still says “users,” “traction,” or “later,” the sheet is not written. A population you do not belong to is a quote, not a prior.
What does not count as this sheet
Not this sheet: a Yibud overall score, a CB Insights reason percent treated as your probability, “90% of startups fail” with no dataset, an establishment rate quoted as a venture-backed rate, or survival treated as success.
You may still read those pages. You may not let them fill in a blank you refused to write.
Not the same page
Base rate vs your risks, your contracts, your pre-mortem, and a score
Several Yibud pages sit next to this one. Mixing them up turns a population table into a personal verdict — or a score into a survival probability.
Risk assessment — your idea’s risks, not the population rate
That page ranks the assumptions your one-liner carries. This page is the outside-view number for a named census. A high-risk idea in a population with a 50% five-year survival rate is still a high-risk idea.
Startup risk assessment →Kill, pivot, decision — contracts that need a definition first
Those pages write Stop, rewrite a dimension, or read a week of evidence. This page is why you must define failure in your own terms before those contracts mean anything. A kill line without a definition is a mood with a date.
Pre-mortem — your specific failure, not the outside view
A pre-mortem invents the history of how this one-liner already died. This page gives the population number you can hold next to that history. The table is not a substitute for writing Rank 1.
Startup idea pre-mortem →Critical assumption — what to test first
Once the definition and the kill line exist, that page is the cheap week against one claim. This page does not run the week. It tells you why “test the riskiest assumption first” is the useful move after you have seen the base rate.
Critical assumption first →Yibud scores — not a survival probability
The validation-score and score-calculator pages explain Startup MRI numbers from a deterministic engine on a one-line idea plus five questions. Those scores are flashlights on assumptions. They are not the chance you will still be operating in March five years from now. Do not read a mid-band score as a percent chance you will fail.
Worked example
One failure-definition sheet you can copy (illustrative)
The week below is illustrative — names and the product are invented so you can see the setup. It is not a study, not a Yibud report, and not a claim that this founder matches any BLS cohort. Copy the method, not the story. The percents in the setup are the Table 7 and Ghosh figures printed above.
The bet, written before the first build week
Illustrative names: founder Nora Patel at Shiftlist, a scheduling waitlist for independent yoga studios. She has not hired and has not raised venture money. Tempting slogan: “90% of startups fail, so I should either quit or go huge.” Useful sheet: she does not belong to Table 7 (no payroll) and she does not belong to Ghosh’s sample (no venture check).
What she writes — population, definition, kill line, assumption
Population: no-payroll side project. Failure definition: I spend six months and still have fewer than three studios in the named city paying a $29/month waitlist. Kill line: 1 June — fewer than three paying studios → Kill the current one-liner; three to five and mixed use → Pivot the person or the job; six or more paying and using it weekly → Continue to the next cheapest test. Riskiest assumption first: a named studio owner will pay $29/month to replace a shared spreadsheet this month, not “someday.”
How she uses the outside view without stealing a census
She reads Table 7 as: new U.S. employer establishments often thin out early — about 1 in 5 gone in a year, about half gone in five. That is a reason to date the kill line, not a reason to write “I am a 1994 establishment.” She reads Ghosh as: if she ever raises venture money, “failure” will mean whatever definition her investors use, and that number can sit anywhere from 30–40% to 90–95%. She does not average those figures into a personal destiny.
What she refuses to do with the slogan and the score
She does not paste “90% fail” into the deck. She does not treat a mid-60s Yibud score as a survival probability. If she later runs the free analyzer, she uses the weakest dimension to pick which assumption the first test should target — then she honors the 1 June line she already wrote.
Write the four fields before the first build week. Use the government table and the venture ladder as priors for the populations they actually measure. Honor the kill line you dated. The method is the sheet, not the story you tell after.
Where Yibud fits
The base rate is the outside view; the report names which assumption to test first
Yibud is a free, no-signup startup idea validator. Scores come from a deterministic rule engine, not from a language model guessing success. Optional AI text, when it is used, only polishes prose. It does not invent the number. A Startup MRI report can name a weak demand, distribution, or monetization dimension. That is a flashlight on which assumption to test first — not a survival probability, not a BLS cohort, and not Ghosh’s ladder. The base rate stays the outside view. If you already have a report, start the sheet with the named assumption, not with the overall score. The method on this page stands alone if you never open the analyzer.
Analyze my idea →Common mistakes
What usually wastes a failure-rate quote
Quoting “90% of startups fail” with no dataset
The sentence is usually repeated without a primary source. This page does not attribute it, and it does not claim a named study has debunked it. If you cannot open the table, do not put the percent in a deck.
Mixing employer establishments with venture-backed startups
Table 7 is new U.S. private-sector employer establishments. Ghosh’s ladder is venture-backed firms, as reported by HBS and the Journal. Averaging 50% five-year survival with “three-quarters did not return capital” produces a number nobody measured.
Treating survival as success — or a closure as failure
BLS “operating” means positive employment. It says nothing about profit or growth. A founder who hits a pre-written kill line and stops has not “failed” in the sense that matters for the next bet. A zombie location with payroll has not “succeeded.”
Treating the base rate as your fate
About half of new employer establishments are gone by year five. That is a prior about a census. It is not a forecast that you will be gone on a Thursday. Write the kill line; collect evidence; update. Do not outsource the decision to a headline.
Using a reasons list as a probability
CB Insights’ 70 / 43 / 29 / 19 percent figures count reasons among 431 shutdowns (385 with identifiable reasons). They are pattern frequency among companies that already died. They are not the chance your idea dies of poor fit.
Sources
Where these numbers come from
- U.S. Bureau of Labor Statistics, Business Employment Dynamics, Table 7. Survival of private sector establishments by opening year (Total private text table) — Used for every cohort percent and count on this page (1994, 2014, 2015, 2020, 2024 examples; 1-, 5-, and 10-year ranges; 1994 20-year and March 2025 readings). Survival rate = share of the opening-year cohort still operating (positive employment) in a later March. Data run through March 2025. This page does not re-derive or add industry splits.
- U.S. Bureau of Labor Statistics, Business Employment Dynamics — Business employment by age and survival of establishments (index) — Index page for the age and survival tables, including Table 7. Used as the official entry point next to the text table.
- U.S. Bureau of Labor Statistics, Business Employment Dynamics frequently asked questions — Used for the caveats on this page: QCEW administrative census (employers only; zero-employment and self-employed without payroll excluded); establishment versus firm; U.S. private sector; “operating” as positive employment, not profit or growth.
- Harvard Business School Working Knowledge, “Why Companies Fail—and How Their Founders Can Bounce Back” (7 March 2011) — Used for Ghosh’s definitions ladder as reported there: 30–40% if failure is liquidation with investors losing most or all of their money; 70–80% if failure is not seeing the projected return on investment; 90–95% if failure is declaring a projection and falling short. Labeled as one researcher’s estimates reported by the press, not an official dataset.
- Deborah Gage, “The Venture Capital Secret: 3 Out of 4 Start-Ups Fail,” Wall Street Journal (20 September 2012) — Used for Ghosh’s finding that about three-quarters of U.S. venture-backed firms do not return investors’ capital (more than 2,000 companies; generally at least $1 million; 2004–2010); for the NVCA estimate of 25% to 30% of venture-backed businesses failing; and for Ghosh’s timing claim that non-venture-backed companies fail more often in the first four years while venture-backed ones tend to fail after year four once investors stop injecting capital. Labeled as press-reported estimates.
- CB Insights, “The top 9 reasons startups fail” (report analyzing 431 VC-backed shutdowns since 2023; updated March 5, 2026) — Used only as already cited elsewhere in this repository: 431 shutdowns; 385 with identifiable reasons; 70 / 43 / 29 / 19 percent as pattern frequency among companies that died, not as a probability of dying. This page invents no other percentages.
In one paragraph
Summary you can quote
“What percentage of startups fail?” depends on the population and on the definition. In BLS Business Employment Dynamics Table 7 (U.S. private-sector employer establishments; through March 2025), about 1 in 5 is gone within a year, roughly half survive five years, and about a third survive ten — and those bands barely move across thirty years (1-year 75.2–80.9%; 5-year 45.4–51.9%; 10-year 32.4–35.3%). The table is employers only, establishments not firms, U.S. private only; operating means positive employment, not profit. Venture-backed “failure,” as reported from Shikhar Ghosh (HBS Working Knowledge 2011; WSJ 2012), runs 30–40% for liquidation, about three-quarters for not returning capital, 70–80% for missing projected ROI, and 90–95% for missing a stated projection. “90% of startups fail” is usually quoted without a primary source. CB Insights’ 70 / 43 / 29 / 19 figures are reason frequencies among shutdowns, not a death probability. A base rate is an outside view, not your fate. A Yibud score is not a survival probability. Write your population, failure definition, kill line and date, and riskiest assumption first.
FAQ
Questions founders actually ask
What percentage of startups fail?
It depends on who you count and what you call failure. For new U.S. private-sector employer establishments, BLS Table 7 shows about 1 in 5 gone within a year, roughly half still operating at five years, and about a third at ten. For venture-backed firms, Ghosh’s reported ladder runs from 30–40% (liquidation) to 90–95% (missed a stated projection), with about three-quarters not returning capital in the Journal write-up. A no-payroll side project is in neither census.
Is it true that 90% of startups fail?
That sentence is usually quoted without a primary source. This page does not attribute it to anyone, and it does not claim a named study has debunked it. A 90–95% figure does appear on Ghosh’s ladder — if failure means declaring a projection and falling short. That is not the same claim as “nine in ten new businesses close.”
If a business is still operating, did it succeed?
Not on this page. BLS survival means positive employment in a later March. It says nothing about profit or growth. A closure is not always a failure either — a founder can hit a pre-written kill line and stop. Write your own definition before you use either word.
Is a Yibud score a survival probability? Do I need an account?
No, and no. Yibud scores come from a deterministic rule engine on your idea inputs. They rank relative risk across dimensions. They are not the chance you will still be operating in five years. You do not need an account. The method on this page stands alone. Run the free analyzer if you want a flashlight on which assumption to test first. Optional language-model text only polishes prose.
How should I use a base rate before I start?
As an outside view, then write four fields: your population, your failure definition (money, time, or a signal), a kill line and date, and the riskiest assumption you will test first. Use Table 7 only if you are talking about U.S. employer establishments. Use Ghosh only if you mean venture-backed outcomes under a named definition. Do not treat either number as your fate.
Write the failure definition before the first build week
The base rate is the outside view. A Yibud report’s weakest dimension tells you which assumption to test first. The kill line is still yours to date.
Generate a Yibud report →